[ad_1]
Discussion
We combined predictive modelling and clustering to identify discrete subgroups of patients with asthma at higher risk of severe AEs.
A history of AE encounters in acute care settings is a known predictor of similar encounters.24 In our study, a 24-month history of all-cause ER visits was the most important predictor. We also observed a 24-month history of all-cause acute care visits in over 90% of patients in clusters 1–4. Among these highly burdened groups, clusters 1 and 4 had AE claims in each patient’s lookback. By contrast, clusters 2 and 3 had AE claims in no more than 3% of patients’ lookback. The latter two had fewer claims for oral steroids as well as SABAs and ICS within 30 days of each other. This may reflect underprescribing, potentially because of fewer historical AE encounters; underfilling prescriptions, potentially reflecting non-adherence; or a lower need for SABAs, suggesting better controlled asthma. The latter is less likely, as patients across all clusters were positive for future AEs. Additionally, non-adherence to ICS may be more common in patients with anxiety and depression, which was more common in cluster 2.8 25–27 Overall, AEs and oral steroids are among the most important predictors. Their low frequency in clusters 2 and 3 explains some of these subgroups’ poor ML model performance, and their higher frequency in clusters 1 and 4 partially explains these subgroups’ best performance.
Cluster 5 was the lowest burden group, with significantly lower frequencies of AEs, all-cause acute care and ER visits. This is a salient finding, as these patients could be easily overlooked during risk assessments. Cluster 5 had significantly more White, younger patients. It has significantly fewer claims for first-hand tobacco exposure, all-cause and AE-related encounters, and prescription claims for oral corticosteroids. These characteristics are the opposite of those expected in high-risk patients.6 Accordingly, whether these patients face risks of AEs in the next 3 months that are more likely to be overlooked in clinical settings warrants further investigation.
We observed a higher frequency of ‘other/unspecified asthma’ in clusters 1 and 4. Both clusters have the highest incidence of AE-related care encounters and the most prescription claims for SABAs and oral steroids. It is unclear whether these findings are coincidental or correlated. Non-specific coding reflects deficiencies in clinician documentation, downstream coding or both. One explanation may be that assessing and documenting asthma control is potentially less likely to occur in time-constrained acute care settings. Non-specific documentation of asthma control may also reflect failures among clinicians to accurately assess severity, potentially increasing the risk of suboptimal management, of more severe AEs, and thereby triggering more SABAs and oral steroid prescriptions. Further studies are warranted to assess whether incomplete documentation of asthma severity correlates with the risk of severe AEs.
Our predictive model achieved a high specificity and NPV, and both metrics were robust to changes in sensitivity. By contrast, as is often the case with healthcare datasets, which are highly imbalanced, a steep trade-off exists between the model’s sensitivity and PPV across different thresholds for positivity. One way to navigate this is to run models at a higher threshold for positivity in series with cluster analyses. This enables a more comprehensive identification and characterisation of patients afflicted by the outcome of interest. In our analysis, patients from clusters 2, 3 and 5 had lower model scores yet represented 58.5% of patients experiencing the outcome. Without our cluster analysis, many of these patients would have been unaccounted for when characterising highly burdened patients.
Prior cluster analyses include survey data for fields unavailable in claims data (ie, results of the asthma control test), making direct comparisons to our analysis difficult.14–16 Despite this, our findings align with those of Sekiya et al, which describe a cluster with the highest frequency of unscheduled asthma-related encounters and treatment with oral corticosteroids and SABAs (like our cluster 1); a second cluster comprised of older females (like our cluster 2) and a third characterised by more frequent intermittent asthma, lower adoption of ICS and fewer asthma hospitalisations.15 The latter is similar to our clusters 2 and 3; among their claims specifying asthma severity, the most common were for intermittent asthma, and both clusters displayed fewer ICS claims. Using EHR data, Wu et al detected a young exacerbator patient phenotype with a high rate of AE recurrence (like our cluster 4),16 which could be due to lower adherence among younger patients with asthma.28 Finally, Zein et al analysed EHR data from 60 302 patients using data not readily available in claims (ie, IgE levels).29 Despite the latter, their models achieved AUROCs comparable to ours.
Our analysis has several advantages. First, we analysed observational data from a large cohort of patients with all severities represented, potentially making it highly generalisable to the broader asthma population in the USA. Second, our claims-based variables are readily accessible to specialists and generalists, which may complement known predictors of asthma AE risk more available in a specialty care setting, such as eosinophil counts and FeNO measurements.13 Finally, we paired our ML model with a cluster analysis that achieved stable cluster performance across different independent data subsets.
Our analysis has several limitations. Our model lacks data on biomarkers such as eosinophil counts and FeNO, which are known to be highly predictive of AEs.13 Second, we did not explicitly calculate adherence metrics for ICS or quantify SABA overuse, which are known to influence asthma mortality and AE risk.30 Third, we selected DD features with large absolute differences in prevalence between positive and negative patients. While this may have introduced bias against low-prevalence features of large effects, our model also included KD features that may have been less prevalent. Fourth, when analysing claims for oral steroids, we did not distinguish between short-term and chronic use. This is important, as the latter applies to patients who are likely at greater risk of our modelled outcome. Finally, claims data have limited internal validity and may underrepresent the prevalence of key risk factors. This may increase the risk of type I and II epidemiological errors.
[ad_2]
Source link
